When drought goes underground: a machine learning approach to groundwater drought

Groundwater drought is a major risk in Mediterranean coastal aquifers, where climate change and intensive abstraction intensify this risk. Despite its impacts on water supply, agriculture, and ecosystems, groundwater drought remains poorly quantified because monitoring networks are sparse and other traditional methods require extensive effort, data and high computational effort. This study develops a data-driven framework to detect, characterize, and map groundwater drought using hydroclimatic predictors and observed groundwater levels. A random forest classifier trained on BIGBANG regional dataset variables is applied to the Bruna River catchment (southern Tuscany, Italy), with drought events defined using a Q20 groundwater threshold. Long-term aggregated meteorological indices (SPEI9 and SPI12) dominate groundwater drought prediction, and the full predictor set yields the best classification performance with clear spatial heterogeneity. The results show that groundwater drought can be reliably detected using a limited set of well-selected drought indicators within a machine-learning framework. This approach enables transferable, operational groundwater drought monitoring and supports proactive groundwater management under intensifying Mediterranean warming and drying trends.